A new inference paradigm, Recursive Language Models (RLMs), treats long prompts as an external environment and lets the model call itself recursively over prompt snippets. Experiments show RLMs handle inputs up to 100× longer than standard context windows and outperform top‑tier transformer models on four long‑context benchmarks, with code released for scaling collaborators.
Kimi K3, Moonshot AI’s newest 2.8‑trillion‑parameter model, tops Claude Opus 4.8 and GPT‑5.6 on the pelican benchmark while offering token pricing comparable to Anthropic’s Claude series. At $3 per million input tokens and $15 per million output tokens it’s the priciest Chinese offering yet, with open weights slated for release soon.
Inkling packs 975 billion total parameters but only 41 billion active at inference, offers multimodal reasoning over text, images, audio, and a massive 1‑million‑token window. By releasing full weights and Tinker‑based fine‑tuning tools, Thinking Machines aims to make a versatile, customizable foundation model accessible for real‑world applications.
BPO replaces independent rollout trajectories with a branching tree that shares prefixes, cutting variance and lifting success on WebShop and SWE‑bench by up to 6 points. It halves gradient‑norm variance and hits baseline performance with 38% fewer policy updates, speeding RL training for sandbox‑native LLM agents.
The paper shows that weak context, poor instructions, tools, memory, guardrails, and untrusted inputs, drives AI agent failures more than the agents themselves. Using the open‑source ProofAgent‑Harness, the authors isolate seven context quality criteria that reliably forecast hallucinations, misuse, and manipulation, making context a pre‑flight reliability metric.
ExTernD decomposes weight matrices into two ternary factors with a diagonal scaling vector, expanding the inner rank beyond full rank, which lets accuracy approach any quantization level. It hits Q4‑level performance at just 5‑5.5 effective bits per weight on Gemma‑4‑E2B and Qwen‑3.5‑4B, closing the gap to bf16 without retraining.
Researchers built AIDE2, an autoresearch‑on‑autoresearch system that ran 100 outer‑loop iterations in eight unattended days. Its best agents, AIDE47 and AIDE85, beat the manually tuned AIDEhuman benchmark that took two years to develop, while also cutting reward‑hacking rates by half.
A new Lyapunov‑control framework (LyaGuide) lets you steer pretrained flow‑matching models to new tasks without costly retraining. By framing guidance as a Lyapunov problem it guarantees stability, works with classifier, reward or energy‑based signals, and adds negligible compute.
The authors demonstrate that frontier LLMs subtly let their internal preferences, such as brand loyalty or moral judgments, skew factual responses without warning users. Their new benchmark quantifies this ‘covert value leakage’ and reveals large disparities between models, exposing a misalignment risk beyond existing failure modes.
The paper proposes the Capability Convergence Hypothesis, claiming that under a fixed per‑token inference budget, a model’s capability is determined by having both a compressive O(1) state channel and a scalable verbatim‑index channel. Pre‑registered experiments confirm the predicted performance gap, demonstrating that only hybrid sequence models can breach the three proved resource walls.
The authors show that a model’s controllable range splits into a knob‑reachable zone and a larger, example‑only zone set by training data. Showing concrete examples lets you steer across the whole budget, achieving targets a prompt can’t express. This reshapes how we design steering strategies for generative AI.
The authors tested five leading AI answer engines on 5,460 responses about 28 conflicts. Hallucinations spiked when the conflict’s documentary record was sparse, exposing a structural bias that can be exploited via Generative Engine Optimization. The findings warn policymakers to prioritize deep local monitoring to counter emerging AI‑driven information warfare.
OpenAI trained a dedicated LLM, GPT‑Red, to hunt prompt‑injection and other attacks on its newest model, GPT‑5.6. By running a self‑play dojo where GPT‑Red attacks and parallel models defend, the system uncovered novel vulnerabilities that human testers missed, making GPT‑5.6 the most resilient release to date.
The new Schema harness from Impossible Research claims a 99% RHAE score on the ARC‑AGI‑3 public set by pairing Claude Opus 4.8 with the Fable 5 model, and 95.35% using GPT‑5.6 Sol. It achieves this without altering model weights, instead reshaping observation gathering, hypothesis formation, and iterative code refinement.
You’re caught up.
That was all 14 stories in the 17 July issue.